feat(mars): add native Mars operator specializations

[Ceng23333]

Summary

• Add Mars glue under src/native/cuda/mars/ (blas.h, blas_utils.h, caster.cuh, runtime_*.h, device_*.h, device_property.h).
• Add thin Operator<…, Device::Type::kMars> specializations under src/native/cuda/mars/ops/*/kernel.h, mirroring src/native/cuda/metax/ and reusing shared src/native/cuda/ops kernels.
• Add Mars GEMM HCBLAS entry src/native/cuda/mars/ops/gemm/hcblas.h.
• No CMake/htcc or torch/smoke changes in this PR (see feat(mars): add WITH_MARS platform and htcc wrapper InfiniTensor/InfiniOps#941 / fix(mars): enable torch bindings and CONTRIBUTING smoke InfiniTensor/InfiniOps#943).

Motivation

Isolate Mars kernel / operator review from platform toolchain and torch bindings. Layout policy matches MetaX: vendor tree under src/native/cuda/mars/, shared CUDA ops under src/native/cuda/ops.

Depends on InfiniTensor#941 (feat/mars-platform-htcc).

Type of Change

• feat — new feature / new operator / new platform
• fix — bug fix
• perf — performance improvement (no behavioral change)
• refactor — code restructuring without behavior change
• test — adding or fixing tests only
• docs — documentation only
• build / ci — build system or CI configuration
• chore — tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Note: Mars / HPCC (WITH_MARS) — native operator specializations only. Template has no Mars checkbox yet.

Smoke Test Result

Requires InfiniTensor#941 merged or stacked head. Host: MetaX X203 / HPCC.

cmake -S . -B build-mars-smoke \
  -DWITH_CPU=ON -DWITH_MARS=ON \
  -DAUTO_DETECT_DEVICES=OFF -DAUTO_DETECT_BACKENDS=OFF \
  -DGENERATE_PYTHON_BINDINGS=OFF \
  -DINFINI_OPS_SMOKE_BUILD=ON \
  -DINFINI_RT_ROOT=<mars-rt-prefix> \
  -DCMAKE_BUILD_TYPE=Release
cmake --build build-mars-smoke --target infiniops --parallel $(nproc)
# Built target infiniops (Mars Operator<…, kMars> specializations linked)

Full pytest -m smoke is on InfiniTensor#943.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	no	N/A - not affected
Iluvatar	no	N/A - not affected
MetaX	no	N/A - not affected
Cambricon	no	N/A - not affected
Moore	no	N/A - not affected
Ascend	no	N/A - not affected
Mars/HPCC (WITH_MARS)	yes	library smoke build passed (infiniops, INFINI_OPS_SMOKE_BUILD=ON)	Pytest smoke on InfiniTensor#943. Smoke allowlist ops without Mars ports (e.g. cast/cat/linear/matmul) remain skip-until-ported.

Full `pytest` output (optional)

N/A — pytest smoke is on stack tip #943.

Benchmark / Performance Impact

N/A

Notes for Reviewers

• Depends on feat(mars): add WITH_MARS platform and htcc wrapper InfiniTensor/InfiniOps#941. Stack: feat(mars): add WITH_MARS platform and htcc wrapper InfiniTensor/InfiniOps#941 → this PR (feat(mars): add native Mars operator specializations InfiniTensor/InfiniOps#942) → fix(mars): enable torch bindings and CONTRIBUTING smoke InfiniTensor/InfiniOps#943.
• Kernel-only diff vs platform branch: feat/mars-platform-htcc...feat/mars-native-kernels
  Fork stacked PR (base feat/mars-platform-htcc): feat(mars): add native Mars operator specializations #2
• Do not rewrite Mars headers to #include metax (kMars vs kMetax).
• After feat(mars): add WITH_MARS platform and htcc wrapper InfiniTensor/InfiniOps#941 merges, retarget this PR base to master.

[Ceng23333]

Closing fork-local stacked PR.

Use InfiniTensor upstream PRs only:

• feat(mars): add WITH_MARS platform and htcc wrapper InfiniTensor/InfiniOps#941 (platform)
• feat(mars): add native Mars operator specializations InfiniTensor/InfiniOps#942 (kernels)
• fix(mars): enable torch bindings and CONTRIBUTING smoke InfiniTensor/InfiniOps#943 (torch/smoke)